Adversarial Weakly Supervised Domain Adaptation for Few Shot Sentiment Analysis

The ability of deep neural networks to generate state-of-the-art results on many NLP problems has been apparent to everyone for some years now. However, when there is not enough labeled data or the test dataset has domain shift, these networks face many challenges and results are getting worse.In this article, we present a method for adapting the domain from formal to colloquial (in sentiment classification). Our method uses two approaches, adversarial training and weak supervision, and only needs a few shots of labeled data.In the first stage, we label a crawled dataset (containing colloquial and formal sentences) with weakly supervised sentiment tags using a sentiment vocabulary network. Then we fine-tune a pre-trained model with adversarial training on this weak dataset to generate domain-independent representations. In the last stage, we train the above fine-tuned neural network with just 50 samples of data (formal domain) and test it on colloquial.Experimental results show that our method outperforms the state-of-the-art model (Pars BERT) on the same data with 15% higher F1 measure.